Research map / Architectures and efficiency / Pruning and efficient tuning
Mitigating Error Propagation in Low-Rank Approximation of Large Models via Distribution-Aware Whitening
Yao Zhu, Rui Sun, Xiu Yan, Yang Li, Wang Lu, Yunjian Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| re-estimate whitening from compressed-layer inputs (distribution-aware) instead of static whitening | SVD-LLM style compression · WikiText-2, ratio 0.3 | perplexity: 33.176 -> 23.702 on LLaMA3-8B | Improved |
| distribution-aware whitening on LLaMA-13B at ratio 0.3 on WikiText-2 | SVD-LLM style compression · WikiText-2, ratio 0.3, LLaMA-13B | perplexity: 7.618 -> 7.635 | No clear effect |
| distribution-aware whitening at higher compression ratio | SVD-LLM style compression · WikiText-2, ratio 0.5, LLaMA3-8B | perplexity: 3908.397 -> 83.950 | Improved |
| apply distribution-aware compression before GPTQ quantization | compressed Vicuna-v1.5-7B · C4, GPTQ-4bit | perplexity: SVD-LLM 85.03, 34.77 worse than ours | Improved |
| apply distribution-aware compression before GPTQ 8-bit | compressed Vicuna-v1.5-7B · C4, GPTQ-8bit | perplexity: SVD-LLM only 0.31 higher | No clear effect |
| add post-compression fine-tuning (full update) on top of distribution-aware compression | compressed LLaMA3-8B · PTB, ratio 0.5 | perplexity: 128.2454 -> 64.5531 vs SVD-LLM | Improved |
| use distribution-aware vs PiSSA initialization for LoRA | LoRA fine-tuning, Vicuna-V1.5-7B · GSM8K, rank 256 | accuracy: 0.320 -> 0.384 | Improved |
| use distribution-aware vs PiSSA initialization for LoRA | LoRA fine-tuning, Qwen2-7B · Object Counting, rank 256 | accuracy: 0.680 -> 0.848 | Improved |
| use distribution-aware vs PiSSA initialization for LoRA at rank 64 | LoRA fine-tuning, Vicuna-V1.5-7B · HumanEval and HumanEval+ | pass rate: 0.183 -> 0.183; 0.165 -> 0.165 | No clear effect |
| use distribution-aware init instead of Kaiming init for LoRA | LoRA fine-tuning, Vicuna-V1.5-7B · MetaMathQA / Code Feedback training loss | training loss | Improved |
| compare against pruning methods (FLAP, LLM-Pruner) after fine-tuning | Vicuna-v1.5-7B compressed · C4 | perplexity | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 9 of 10 shared results
- PiSSA — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.87)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.87)
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · similarity 0.86)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.86)
- Determining Layer-wise Sparsity for Large Language Models Through a Theoretical Perspective (ICML 2025 · similarity 0.86)
- LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
